Your Brand Now Has a Second Reader, and It Can't See Figma

Ask a coding agent to build a screen today and watch what it grabs. A gray Material button. A shadcn card. A stack of defaults that belong to no brand at all. That's not the agent failing you. It's the agent working from nothing, because your brand lives in Figma files and the heads of three senior designers, and a model can't read any of that at build time.

Something shifted in who consumes your work first. For years the answer was a person. Now it's often a machine, and the machine can't see most of what you built. Let me catch you up on what that means and what to do about it.

The file that speaks both languages

The fix making the rounds is one plain-text file, checked into your repository. It states your design two ways at once: machine-readable tokens up top, human-readable reasoning below. Google Labs open-sourced the DESIGN.md format in April 2026, pitching it as an agent standard that sits next to README.md at the root.

The tokens give an agent your exact colors, type scale, and corner radius. The prose gives it judgment: why the primary is a deep forest green, when a designer is allowed to break the grid. Ask for a button with no file in play and you get a blue-gray rectangle, correct and anonymous. Ask with the file present and you get your green at your radius.

The prose half is the part teams skip, and it carries the weight. A stylesheet already holds your hex values. What it does not hold is the reasoning. Skip the prose and you have a palette. Write it and you have a brief.

Cheaper to start than you think

You do not need a repository or a committee. Open whatever assistant you already use, hand it an example file, and ask it to interview you: your core colors and what each should signal, two or three type levels, the radius that feels right. Then have it format your answers and paste them into a plain-text file. That draft took a conversation.

Name tokens by role, not by look. A token called bigRedButton breaks the moment you rebrand. A token called button.primary survives any visual change, because the role outlasts the color. This is Brad Frost's Atomic Design argument wearing new clothes: a shared component language beats a private one an agent has to guess at.

The format ships tooling that lints and diffs. The linter flags color pairs that fail WCAG contrast, catching a review that used to happen by eye. Point an assistant at your file and your live site and ask where they disagree. A source of truth nobody reconciles turns into fiction.

The structure underneath the pixels

DESIGN.md handles how your brand looks. The bigger problem sits one layer down, in how your content is organized. Information architecture was the discipline nobody funded for twenty years, because its payoff was "fewer confused users," a number no one could put on a slide.

That changed when poor structure started showing up on the balance sheet. The same gap that once cost you a confused visitor now costs you a hallucination, an agent that confidently retrieves garbage and acts on it. In a 2025 survey, data quality and availability topped the list of AI adoption barriers, above every other obstacle. Retrieval doesn't find the right document. Without structure, it finds the loudest one, the one that shares the most words with the query whether or not it's true or current.

People forgive a messy label. They use judgment, work around it, move on. A model removes the human from that loop. It pattern-matches over the mess and reproduces it at scale, confidently. The mess didn't get worse. The blast radius did.

When wrong answers start moving money

Answering a question wrong is embarrassing. Taking an action wrong is a liability. The moment an agent stops retrieving and starts doing, routing a ticket, updating a record, approving a request, it needs to know what belongs to what and what it's allowed to touch.

There's already a price tag. In Moffatt v. Air Canada, the airline's chatbot told a grieving customer he could claim a bereavement fare retroactively, the opposite of what the airline's own policy page said. The bot even linked to the page that contradicted it. A tribunal held the airline liable. Two pages said opposite things, and nothing reconciled them.

The upside is measurable too. Anthropic's Contextual Retrieval work cut failed retrievals by up to 49 percent, a reliability gain from fixing content, not swapping models. You can buy a sharper model and still serve wrong answers, because the model retrieves from a pile nobody agreed how to organize.

The surface you no longer own

Here's the part that reframes the whole job. For a growing set of tasks, your user never opens your app. Aurélie Radom puts it plainly: AI changed who reads your work first. A user asks Claude to find every auto-renewal clause in their contracts, Claude reaches your platform through a connector, does the work, and answers in chat. They never see your navigation, your dashboard, or your brand.

When work flows through an assistant, the interface your user sees is the assistant's, not yours. Your product did the job and stayed invisible. That's a harder design problem, not an easier one. Your job shifts from designing the destination to being legible through someone else's surface.

Which loops back to why the machine-readable file matters. Generation is cheap now. A convincing interface takes seconds. But a screenshot hides the spacing system, the component behavior, the accessibility calls, everything that makes a product survive its first release. The easier interfaces get to generate, the more the invisible layer is worth.

The deep cut

Don't treat DESIGN.md and information architecture as two projects. They're the same move: writing down what used to live in people's heads so a machine can read it. The trap is doing the visible half and skipping the structural half. A gorgeous machine-readable brand file sitting on top of a content pile nobody organized still ships wrong answers, just in your exact forest green.

And retire the metrics that reward time-in-app. If the assistant does the task, your page views and click counts fall even when you're winning. Judge yourself on whether the agent got it right, not on whether the user visited you to watch it happen.

Three questions for your team

1. If a coding agent built a screen from our repository right now, with no designer in the room, would it produce our brand or a generic default? If the answer is default, your visual identity isn't legible to what builds it.

2. Where in our content do two pages say opposite things? Point an assistant at your live product and ask it, because that's the Air Canada failure waiting to happen, and it's cheaper to find now than in a tribunal.

3. Which of our tasks already flow through an assistant our user never leaves? For those, we don't own the surface anymore, and we should be measuring whether the agent got it right, not whether anyone opened our app.